eduKateSG · Why Science?
Look beyond a shift in the average and ask whether a gene changes its modes, mixture proportions or expressed-state means across two conditions
Reading routes
Science learning becomes useful when a familiar object or observation is turned into a system of quantities, mechanisms and claim limits. This guide owns one applied evidence-reading job inside eduKateSG’s wider Science estate. It connects naturally to Why Science Single Cell Rna Sequencing Barcodes Transcriptome Heterogeneity Evidence; Why Science Mast Hurdle Models Single Cell Differential Expression; Why Science Muscat Multi Sample Multi Group Differential State Analysis; Why Science Sccomp Differential Composition Variability Single Cell Data; Education Hub; Singapore Secondary School Directory; Career Adulthood Hub; Science Learning Hub. It also keeps current school and public claims traceable to visible primary sources: scDD primary study; scDD PubMed record; scDD open full text; Official Bioconductor scDD package; 2026 Singapore–Cambridge O-Level Biology syllabus; MOE G2/G3 Lower Secondary Science syllabus; 2026 MOE G2 Computing syllabus. The sources describe the scientific scope; this article translates that scope into a calm route for Primary Science, PSLE Science, Secondary Science, O-Level Science, STEM exploration, school choices and career pathways without inventing admission or employment outcomes.
scDD, presented by Korthauer and colleagues in Genome Biology in 2016, is a Bayesian method for detecting genes whose single-cell expression distributions differ between two biological conditions. Rather than reducing every difference to a change in the mean, it can classify patterns involving differential expression, modality, mixture proportion, or both modality and component means. Those categories are statistical descriptions of observed distributions; they become biological stories only after cell identity, sampling, technical effects and independent evidence have been examined.
Inside this guide
1–12 · Foundations and models
- 1. Begin beyond the average
- 2. Define a distributional question
- 3. Recognise mixture components
- 4. Separate modes from states
- 5. Protect biological replication
- 6. Choose the expression scale
- 7. Read the 2016 evidence in scope
- 8. Build a metadata map
- 9. Filter low-information genes
- 10. Inspect raw and transformed distributions
- 11. Check cell-type coherence
- 12. Understand the Bayesian hierarchy
13–24 · Evidence, testing and applications
- 13. Estimate components without overcounting
- 14. Compare conditions jointly
- 15. Practise with a fictional scDD table
- 16. Understand differential expression
- 17. Understand differential modality
- 18. Understand differential proportion
- 19. Understand combined changes
- 20. Read posterior evidence
- 21. Adjust the discovery rule
- 22. Inspect biological samples separately
- 23. Run transformation sensitivity
- 24. Run cell-subsampling sensitivity
25–36 · Learning, decisions and pathways
- 25. Compare with MAST
- 26. Compare with compositional methods
- 27. Build a distribution evidence card
- 28. Design a validation experiment
- 29. Connect scDD to school science
- 30. Try a same-mean activity
- 31. Connect to mathematics
- 32. Connect to computing and AI literacy
- 33. Connect to school and career pathways
- 34. Create a family evidence habit
- 35. Use precise scDD language
- 36. Finish with a shape-to-mechanism bundle
Section 1 of 36
1. Begin beyond the average
Two groups can have the same mean expression while one has a single broad distribution and the other has low- and high-expression modes. The average hides the structure.
scDD uses flexible Bayesian mixture modelling to compare full single-cell expression distributions rather than relying only on a difference in mean expression.
Archive sample manifests, preprocessing decisions, model formulae, contrasts, random seeds, package versions and exclusions so another analyst can reconstruct the decision path.
Section 2 of 36
2. Define a distributional question
Ask whether the entire expression pattern differs between two conditions, not merely whether one summary statistic moved.
Its reported pattern classes include differential expression, differential modality, differential proportion, and changes involving both modes and component means.
Keep the biological sample as the unit of replication. More cells can sharpen cellular description, but they do not manufacture more independent donors, animals or cultures.
Section 3 of 36
3. Recognise mixture components
A flexible mixture represents subgroups of observations with different locations or shapes. Components are modelling devices that require biological interpretation after fitting.
A mode is a feature of the fitted distribution. It may suggest cell states or mixtures, but it does not prove a distinct biological state without external evidence.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state exactly what false signal each check can reveal.
Section 4 of 36
4. Separate modes from states
A second expression mode may correspond to a cell state, cell subtype, technical artefact or sampling fluctuation. Use markers and experiments to distinguish those explanations.
The original method is framed around two-condition comparisons, so extensions to many groups or complex repeated designs need explicit additional methodology.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful curve is not a complete result.
Section 5 of 36
5. Protect biological replication
Cells from one sample share many influences. Cell-level distributions should be compared with the number and structure of biological samples plainly reported.
Posterior classification depends on prior choices, sample size, expression scale and separation of components; ambiguous genes should remain ambiguous rather than being forced into a story.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 6 of 36
6. Choose the expression scale
Transformation, normalisation and zero handling can change apparent modality. Record the scale and test whether conclusions survive reasonable alternatives.
Distributional change can be informative when averages match, but technical zeros, batch effects and cell-composition shifts can also reshape the observed distribution.
Write association, distributional difference or trajectory-linked pattern when that is what was tested. Reserve cause, diagnosis and benefit for designs that support those stronger claims.
Section 7 of 36
7. Read the 2016 evidence in scope
Korthauer and colleagues introduced scDD as a Bayesian approach for detecting and classifying distributional differences in single-cell expression between two conditions.
scDD uses flexible Bayesian mixture modelling to compare full single-cell expression distributions rather than relying only on a difference in mean expression.
Archive sample manifests, preprocessing decisions, model formulae, contrasts, random seeds, package versions and exclusions so another analyst can reconstruct the decision path.
Section 8 of 36
8. Build a metadata map
Link every cell to sample, condition, batch, cell type and quality metrics. Check whether a supposed distributional change tracks a processing run.
Its reported pattern classes include differential expression, differential modality, differential proportion, and changes involving both modes and component means.
Keep the biological sample as the unit of replication. More cells can sharpen cellular description, but they do not manufacture more independent donors, animals or cultures.
Section 9 of 36
9. Filter low-information genes
Genes with almost no observed expression cannot support complex mixture classification. Apply a declared, outcome-blind information threshold.
A mode is a feature of the fitted distribution. It may suggest cell states or mixtures, but it does not prove a distinct biological state without external evidence.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state exactly what false signal each check can reveal.
Section 10 of 36
10. Inspect raw and transformed distributions
Show per-condition histograms, densities, dot plots and sample-stratified views. Smoothing choices should not create the very modes later tested.
The original method is framed around two-condition comparisons, so extensions to many groups or complex repeated designs need explicit additional methodology.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful curve is not a complete result.
Section 11 of 36
11. Check cell-type coherence
A mixed cluster can produce multiple modes simply because it contains multiple cell types. Revisit annotations and marker patterns before interpreting state transitions.
Posterior classification depends on prior choices, sample size, expression scale and separation of components; ambiguous genes should remain ambiguous rather than being forced into a story.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 12 of 36
12. Understand the Bayesian hierarchy
Prior information and data combine to estimate component structure and differential patterns. Report relevant defaults or changed priors and why they suit the dataset.
Distributional change can be informative when averages match, but technical zeros, batch effects and cell-composition shifts can also reshape the observed distribution.
Write association, distributional difference or trajectory-linked pattern when that is what was tested. Reserve cause, diagnosis and benefit for designs that support those stronger claims.
Section 13 of 36
13. Estimate components without overcounting
More mixture components improve flexibility but can split noise into apparent subgroups. Prefer the simplest distributional description adequately supported by the data.
scDD uses flexible Bayesian mixture modelling to compare full single-cell expression distributions rather than relying only on a difference in mean expression.
Archive sample manifests, preprocessing decisions, model formulae, contrasts, random seeds, package versions and exclusions so another analyst can reconstruct the decision path.
Section 14 of 36
14. Compare conditions jointly
A coherent comparison evaluates how component means and proportions relate across two groups rather than fitting unrelated stories to each group.
Its reported pattern classes include differential expression, differential modality, differential proportion, and changes involving both modes and component means.
Keep the biological sample as the unit of replication. More cells can sharpen cellular description, but they do not manufacture more independent donors, animals or cultures.
Section 15 of 36
15. Practise with a fictional scDD table
This classroom table is invented and is not a result from the scDD paper.
| Fictional gene | Condition A | Condition B | Candidate class | Caution |
|---|---|---|---|---|
| GENE-M | One mode | Two modes | DM | Check mixed cell types |
| GENE-P | Same modes | Different proportions | DP | Check sample balance |
| GENE-E | Shifted mode | Shifted mode | DE | Check transformation |
A mode is a feature of the fitted distribution. It may suggest cell states or mixtures, but it does not prove a distinct biological state without external evidence.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state exactly what false signal each check can reveal.
Section 16 of 36
16. Understand differential expression
In the scDD pattern vocabulary, a gene can show a shift in component mean without changing the number of modes. State the model-specific meaning rather than assuming every DE tool uses identical definitions.
The original method is framed around two-condition comparisons, so extensions to many groups or complex repeated designs need explicit additional methodology.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful curve is not a complete result.
Section 17 of 36
17. Understand differential modality
DM describes a change in the number of modes. It is an invitation to investigate cellular substructure, not a direct count of cell states.
Posterior classification depends on prior choices, sample size, expression scale and separation of components; ambiguous genes should remain ambiguous rather than being forced into a story.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 18 of 36
18. Understand differential proportion
DP describes altered mixing proportions among otherwise aligned components. A compositional shift, sampling imbalance or changed state frequency could contribute.
Distributional change can be informative when averages match, but technical zeros, batch effects and cell-composition shifts can also reshape the observed distribution.
Write association, distributional difference or trajectory-linked pattern when that is what was tested. Reserve cause, diagnosis and benefit for designs that support those stronger claims.
Section 19 of 36
19. Understand combined changes
A gene can differ in both modality and component means. Such patterns are rich but harder to validate, so display the fitted and observed distributions together.
scDD uses flexible Bayesian mixture modelling to compare full single-cell expression distributions rather than relying only on a difference in mean expression.
Archive sample manifests, preprocessing decisions, model formulae, contrasts, random seeds, package versions and exclusions so another analyst can reconstruct the decision path.
Section 20 of 36
20. Read posterior evidence
Posterior probabilities quantify support under the model and priors. They are not universal probabilities that a biological mechanism is true.
Its reported pattern classes include differential expression, differential modality, differential proportion, and changes involving both modes and component means.
Keep the biological sample as the unit of replication. More cells can sharpen cellular description, but they do not manufacture more independent donors, animals or cultures.
Section 21 of 36
21. Adjust the discovery rule
Declare probability thresholds, gene filtering and how false discoveries are estimated or controlled. Avoid choosing cutoffs after seeing the most attractive plots.
A mode is a feature of the fitted distribution. It may suggest cell states or mixtures, but it does not prove a distinct biological state without external evidence.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state exactly what false signal each check can reveal.
Section 22 of 36
22. Inspect biological samples separately
A beautiful bimodal pooled curve may arise because one donor is low and another is high. Overlay or facet by sample to distinguish within-sample modes from between-sample mixing.
The original method is framed around two-condition comparisons, so extensions to many groups or complex repeated designs need explicit additional methodology.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful curve is not a complete result.
Section 23 of 36
23. Run transformation sensitivity
Repeat key genes across defensible normalisation and transformation choices. If modality disappears, report the result as scale-sensitive.
Posterior classification depends on prior choices, sample size, expression scale and separation of components; ambiguous genes should remain ambiguous rather than being forced into a story.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 24 of 36
24. Run cell-subsampling sensitivity
Rebalance cell numbers per sample and condition. A distribution should not earn confidence merely because one sample contributed far more cells.
Distributional change can be informative when averages match, but technical zeros, batch effects and cell-composition shifts can also reshape the observed distribution.
Write association, distributional difference or trajectory-linked pattern when that is what was tested. Reserve cause, diagnosis and benefit for designs that support those stronger claims.
Section 25 of 36
25. Compare with MAST
MAST decomposes detection and positive-expression differences through a hurdle model. scDD asks whether the broader expression distribution and its mixture structure change.
scDD uses flexible Bayesian mixture modelling to compare full single-cell expression distributions rather than relying only on a difference in mean expression.
Archive sample manifests, preprocessing decisions, model formulae, contrasts, random seeds, package versions and exclusions so another analyst can reconstruct the decision path.
Section 26 of 36
26. Compare with compositional methods
A DP classification concerns mixture weights for a gene’s expression components, not automatically the abundance of annotated cell types measured by compositional tools.
Its reported pattern classes include differential expression, differential modality, differential proportion, and changes involving both modes and component means.
Keep the biological sample as the unit of replication. More cells can sharpen cellular description, but they do not manufacture more independent donors, animals or cultures.
Section 27 of 36
27. Build a distribution evidence card
Record gene, cell population, transformation, fitted components, posterior class support, donor coverage, batch checks, sensitivities and biological validation.
A mode is a feature of the fitted distribution. It may suggest cell states or mixtures, but it does not prove a distinct biological state without external evidence.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state exactly what false signal each check can reveal.
Section 28 of 36
28. Design a validation experiment
Use RNA imaging, protein staining, sorted populations or time-course sampling to test whether proposed modes recur and correspond to measurable cell states.
The original method is framed around two-condition comparisons, so extensions to many groups or complex repeated designs need explicit additional methodology.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful curve is not a complete result.
Section 29 of 36
29. Connect scDD to school science
Students learn that two groups can share an average while having different spreads or clusters. Looking at the full distribution produces a fairer comparison.
Posterior classification depends on prior choices, sample size, expression scale and separation of components; ambiguous genes should remain ambiguous rather than being forced into a story.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 30 of 36
30. Try a same-mean activity
Create two sets of numbers with equal means: one tightly centred and one split into low and high groups. Plot them and discuss which summary reveals the difference.
Distributional change can be informative when averages match, but technical zeros, batch effects and cell-composition shifts can also reshape the observed distribution.
Write association, distributional difference or trajectory-linked pattern when that is what was tested. Reserve cause, diagnosis and benefit for designs that support those stronger claims.
Section 31 of 36
31. Connect to mathematics
Mixtures, proportions, probability, modes and posterior evidence show how shapes of data can answer questions that one average cannot.
scDD uses flexible Bayesian mixture modelling to compare full single-cell expression distributions rather than relying only on a difference in mean expression.
Archive sample manifests, preprocessing decisions, model formulae, contrasts, random seeds, package versions and exclusions so another analyst can reconstruct the decision path.
Section 32 of 36
32. Connect to computing and AI literacy
An algorithm can label a pattern DM or DP, but humans must check whether the pattern is stable, technically plausible and biologically meaningful.
Its reported pattern classes include differential expression, differential modality, differential proportion, and changes involving both modes and component means.
Keep the biological sample as the unit of replication. More cells can sharpen cellular description, but they do not manufacture more independent donors, animals or cultures.
Section 33 of 36
33. Connect to school and career pathways
Interest in distributions can lead toward statistics, biology, biotechnology, medicine, data science or computational genomics through varied educational routes.
A mode is a feature of the fitted distribution. It may suggest cell states or mixtures, but it does not prove a distinct biological state without external evidence.
Use negative controls, simulated nulls or label permutations only when their exchangeability assumptions match the design, and state exactly what false signal each check can reveal.
Section 34 of 36
34. Create a family evidence habit
When a report says expression changed, ask whether the average, spread, number of modes or proportions changed and whether individual samples show the same pattern.
The original method is framed around two-condition comparisons, so extensions to many groups or complex repeated designs need explicit additional methodology.
Report effect direction, uncertainty, sample support and sensitivity beside adjusted evidence; a ranked gene list or colourful curve is not a complete result.
Section 35 of 36
35. Use precise scDD language
Say model-supported distributional difference or candidate DM, DP, DE or combined pattern. Avoid declaring a new cell state from a fitted mode alone.
Posterior classification depends on prior choices, sample size, expression scale and separation of components; ambiguous genes should remain ambiguous rather than being forced into a story.
Create an evidence card naming the comparison, measured modality, statistical unit, model, result, validation status, unresolved alternatives and narrowest defensible claim.
Section 36 of 36
36. Finish with a shape-to-mechanism bundle
Deliver metadata, observed plots, transformations, mixture settings, posterior results, sample-stratified checks, sensitivity analyses and a targeted experiment for the proposed modes.
Distributional change can be informative when averages match, but technical zeros, batch effects and cell-composition shifts can also reshape the observed distribution.
Write association, distributional difference or trajectory-linked pattern when that is what was tested. Reserve cause, diagnosis and benefit for designs that support those stronger claims.
## Source date and scope note Primary papers, open records and official software documentation were checked for this article on 10 October 2026. Software interfaces and recommendations can change, so readers should consult the linked current documentation before reproducing an analysis.
## Final reader checklist Before accepting a claim, confirm the biological sample count, condition definition, measured modality, preprocessing, statistical unit, model formula, uncertainty, software version, negative controls, sensitivity checks, independent validation and the exact boundary between association and causation.
Contents · Previous section · Continue to the Science Learning Hub